27 June 2020
Thanks to the increasingly powerful capabilities of machine learning algorithms, artificial intelligence has entered mainstream business. Machine learning algorithms enable computers to train themselves for tasks such as driving a car, controlling a robot, or automating decisions.
But as artificial intelligence begins to handle sensitive tasks, such as helping choose which prisoners are to be released on bail, policy makers have insisted that computer scientists provide assurance that automated systems are designed to avoid or minimize unwanted consequences, such as excessive Risk or racial and gender bias.
A team led by researchers at Stanford University and the University of Massachusetts Amherst published a paper in the journal Science on November 22 some advices. The paper outlines a new technique that can be used to turn ambiguous targets such as avoiding gender bias into precise mathematical standards, so that machine learning algorithms can train artificial intelligence applications to avoid this behavior.
Emma Brunskill, an assistant professor of computer science at Stanford University and senior author of the paper, said: "We want to advance artificial intelligence, respect the values of human users, and justify our trust in autonomous systems."
The premise of this work is that if "unsafe" or "unfair" results or behaviors can be defined mathematically, then it should be possible to create corresponding algorithms that can learn from the data how to avoid unwanted results, and Has a high degree of credibility. Researchers also hope to develop a set of technologies that allow users to specify what behavior constraints they want, so that machine learning designers can safely use systems trained on past data and apply them to real-world environments.
"We show how designers of machine learning algorithms can help other developers, when embedding artificial intelligence into their products and services, they can more easily describe unwanted results or behaviors, and artificial intelligence systems will Avoid these situations with high probability, "said Philip Thomas, assistant professor of computer science at the University of Massachusetts Amherst and the first author of the paper.
Researchers tested their method in an attempt to improve the fairness of the algorithm for predicting college students' GPA based on test scores, a common algorithm that may generate gender bias. They used experimental data sets to provide mathematical instructions to the algorithm to avoid letting the resulting predictive method systematically overestimate or underestimate the GPA of a gender. With these instructions, the algorithm found a better way to predict the GPA of students than existing methods, with much less systematic gender bias. In this regard, previous methods are difficult either because they do not have built-in fairness filters or because the range of algorithms developed to achieve fairness is too limited.
The research team also developed another algorithm and used it to automatically balance the safety and performance of insulin pumps. The pump must decide how much insulin to deliver to the patient at mealtime. Ideally, the pumped insulin just happens to keep blood sugar levels stable. Too little insulin can raise blood sugar, cause short-term discomfort such as nausea, and increase the risk of long-term complications such as cardiovascular disease; excessive use of insulin can cause blood sugar to plummet, which is a potentially fatal consequence.
Machine learning can better assist individuals by identifying subtle patterns in how blood glucose levels respond to different doses of insulin, but existing methods don't make it easy for doctors to determine the results (such as hypoglycemia) that automatic dose algorithms should avoid. Brunskill and Thomas show how to train a pump to determine a dose tailored for a given patient and avoid complications caused by overdosing or underdosing. Although the team isn't ready to test the algorithm on real people, it points out an artificial intelligence approach that could eventually improve the quality of life for people with diabetes.
Brunskill and Thomas used the term "Seldonian algorithm" in their Science paper to define their method, quoting Hari Seldon, a character invented by science fiction author Asimov, who had announced three laws of robotics, which began with It is "the robot should not harm humans, nor should it harm humans by inaction."
Thomas acknowledges that the field still has a long way to go to follow these three laws, but he said that this Seldonian framework will make it easier for machine learning designers to build avoidance instructions into various algorithms, to some extent This allows them to assess the likelihood that the trained system will function properly in the real world.
Brunskill said the proposed framework builds on the efforts of many computer scientists to strike a balance between creating powerful algorithms and developing methods to ensure their reliability.